diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index ef5cafb175..fa9501f58a 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -1084,6 +1084,17 @@ def run_training_process(
                 'Install for better performance: pip install "triton-windows<3.7"'
             )
 
+    # ── 1d. Ensure torch.distributed.is_initialized exists before ML libs load ──
+    # The ROCm Windows wheel (2.9.0+rocmsdk*) does not expose is_initialized on
+    # the torch.distributed module object until it is explicitly imported.
+    # transformers and trl access it before that happens, causing AttributeError.
+    try:
+        import torch.distributed as _td  # noqa: F401 -- forces full module load
+        if not hasattr(_td, "is_initialized"):
+            _td.is_initialized = lambda: False
+    except Exception:
+        pass
+
     # ── 2. Now import ML libraries (fresh in this clean process) ──
     try:
         _send_status(event_queue, "Importing Unsloth...")
diff --git a/studio/backend/utils/wheel_utils.py b/studio/backend/utils/wheel_utils.py
index 3ed9bda827..06c5544f22 100644
--- a/studio/backend/utils/wheel_utils.py
+++ b/studio/backend/utils/wheel_utils.py
@@ -14,6 +14,7 @@ import urllib.request
 from typing import Callable
 
 from utils.native_path_leases import child_env_without_native_path_secret
+from utils.subprocess_compat import windows_hidden_subprocess_kwargs
 
 _logger = logging.getLogger(__name__)
 
@@ -62,6 +63,7 @@ def probe_torch_wheel_env(*, timeout: int | None = None) -> dict[str, str] | Non
             text = True,
             timeout = timeout,
             env = child_env_without_native_path_secret(),
+            **windows_hidden_subprocess_kwargs(),
         )
     except subprocess.TimeoutExpired:
         return None